On Thu, 12 Sep 2024 10:51:23 +0800 Jijie Shao wrote:

+static int hbg_napi_tx_recycle(struct napi_struct *napi, int budget)
+{
+	struct hbg_ring *ring = container_of(napi, struct hbg_ring, napi);
+	/* This smp_load_acquire() pairs with smp_store_release() in
+	 * hbg_start_xmit() called in xmit process.
+	 */
+	u32 ntu = smp_load_acquire(&ring->ntu);
+	struct hbg_priv *priv = ring->priv;
+	struct hbg_buffer *buffer;
+	u32 ntc = ring->ntc;
+	int packet_done = 0;
+
+	while (packet_done < budget) {

you should so some cleanup even if budget is 0
in fact you can hardcode the amount of work Tx NAPI does to 128 and
don't look at the budget as a limit. Per NAPI documentation budget
is for Rx

+		if (unlikely(hbg_queue_is_empty(ntc, ntu, ring)))
+			break;
+
+		/* make sure HW write desc complete */
+		dma_rmb();
+
+		buffer = &ring->queue[ntc];
+		if (buffer->state != HBG_TX_STATE_COMPLETE)
+			break;
+
+		hbg_buffer_free(buffer);
+		ntc = hbg_queue_next_prt(ntc, ring);
+		packet_done++;
+	}
+
+	/* This smp_store_release() pairs with smp_load_acquire() in
+	 * hbg_start_xmit() called in xmit process.
+	 */
+	smp_store_release(&ring->ntc, ntc);
+	netif_wake_queue(priv->netdev);
+
+	if (likely(napi_complete_done(napi, packet_done)))

if packet_done >= budget you should not call napi_complete_done()
as you are not done

+		hbg_hw_irq_enable(priv, HBG_INT_MSK_TX_B, true);